Research On Privacy-preserving Scheme In Regression Learning Problems

In order to ensure the accuracy of the model,traditional machine learning algorithms need to collect a large amount of raw data for model training.On the one hand,people enjoy the convenience brought by high-precision models,such as image processing,text recognition,etc.On the other hand,it also faces the problem of privacy leakage of data,which has aroused people's extensive attention.How to ensure that regression model training is safely carried out without revealing user privacy,or security prediction based on existing models becomes an urgent problem in current machine learning filed.This paper proposes a privacy-preserving regression model based on secure two-party computation for the regression learning problem of how to conduct safety regression model training without revealing user privacy.The model is based on additive secret sharing scheme,which is perfect security in the sense of information theory.In this case of secret sharing,the model uses an auxiliary server to perform the main regression learning calculation process by two noncollusion and semi-honest master computing servers.The main ideas of the full text are:Firstly,by observing the conventional implementation algorithm in the regression learning problem,the two-party computation protocol for the basic operations involved in the privacy implementation of the algorithm is given,such as the addition and multiplication of the secret sharing form.Secondly,considering the inevitable in the protocol real numbers will appear,giving a protocol for safely implementing two numbers comparison in this case;Finally,the regression is given in combination with the above protocol.The privacy implementation form of the conventional algorithm in the regression learning problem is given,and the experimental comparison is made.Experiments show that the privacy-preserving protocol under the privacypreserving model can safely implement all the steps of the conventional algorithm,give consistent calculation results,and the overall solution consumption.At the same time,the privacy of the user's original data is protected,and the privacy of the protection model is also realized.When it is within the acceptable range,it has certain reference significance for studying the privacy realization of other machine learning algorithms.